RESUMO
Forecasting the need for Renal Replacement Therapy (RRT) in intensive care units (ICUs) at an early stage can enhance patient outcomes and optimize resource allocation. The study aimed to develop a model for early prediction of Renal Replacement Therapy (RRT) requirement within 24 hours of ICU admission, utilizing machine learning techniques and SHapley Additive exPlanations (SHAP). It assessed various models including Random Forest (RF), Neural Network (NN), and XGBoost, using data from 34,000 ICU admissions. XGBoost showed superior performance in terms of AUPRC, while RF performed better in AUC-ROC. Results were consistent before and after Principal Component Analysis (PCA) and feature evaluation analysis. The top 10 feature models outperformed the PCA model while using fewer inputs. These findings suggest the potential utility of the developed models in accurately predicting RRT requirement within 24 hours of ICU admission, aiding in efficient critical care delivery.